Casco built an AI whose entire job is to break into your software before someone else does

Most companies find out they have a security hole the same way: a year after the last audit, from someone who was not supposed to be looking. The annual penetration test - that six-week engagement with a boutique firm and an invoice with a comma in it - has been the industry's standard ritual for decades. It produces a PDF. The PDF describes a system that has already changed. Casco looked at that ritual and asked a simple question: why does the attacker only show up once a year?

Casco, a Y Combinator company from the Spring 2025 (X25) batch, sells autonomous security testing. In plain terms, it is an AI that continuously tries to break into your software - your web apps, your APIs, your cloud infrastructure, and, increasingly, the AI systems your team shipped last quarter. When it finds a way in, it writes up what it found, why it matters to the business, and the steps to reproduce it. Then it keeps going. The company operates out of San Francisco under the legal name Transmittr, Inc., and it is small - roughly a dozen people - but the claim it makes is not: Casco says more than 300 companies now run its tests, and that its testing touches AI systems used across a majority of the Fortune 500.

01 / THE FOUNDERSTwo people who spent years securing other people's AI

The company was started in 2025 by Rene Brandel and Ian Saultz. Both came out of Amazon Web Services, where they worked on the Generative AI team - one group building the models and features, and, right next to it, the group responsible for making sure those systems could not be turned against their users. That adjacency matters. Brandel and Saultz did not stumble into AI security from the outside; they watched, from inside one of the largest cloud providers on earth, how quickly AI features ship and how rarely anyone stress-tests them for abuse. Brandel is CEO. Saultz is CTO. Between them they carry background from AWS, Microsoft, and U.S. government security work.

The thesis they left to build is contrarian in a quiet way. Security teams spend most of their budget on defense - firewalls, monitors, alerts, the machinery of waiting for something bad. Casco's product is the opposite posture. It is offense that works for you. The best way to know whether your locks hold, the founders argue, is to hire something that never gets tired of trying every door.

The expertise showing up in the product is specific, not generic. Penetration testing is a craft with a lot of tacit knowledge - the instinct for which parameter looks fragile, the patience to chain three small weaknesses into one real breach, the judgment to tell a scary-looking finding from a genuinely dangerous one. Automating that has historically produced scanners that generate noise: long lists of low-severity flags a human then has to sift. Casco's founders spent their AWS years on exactly the systems where that judgment gets encoded, and the company leans on it as a selling point, promising findings with minimal false positives and clear articulation of business impact rather than a raw dump of alerts.

02 / THE PROBLEMThe pentest was built for software that stopped changing

Traditional penetration testing has a timing problem baked into its business model. A firm scopes an engagement, schedules it weeks out, runs a team of humans against your system for a fixed window, and delivers findings. By the time you read them, you have shipped a dozen releases. The test was a photograph; your software is a film. For a modern company deploying continuously, a once-a-year snapshot is close to useless the moment the ink dries.

Then there is the newer problem, the one that barely existed three years ago. Every company is now shipping AI features - chatbots, agents, workflows that take actions on a user's behalf. Those systems have their own attack surface. They can be jailbroken, tricked into leaking data, or manipulated through patient, multi-step prompting that no traditional scanner was built to catch. Most of the tools security teams already own simply do not look there. And the pressure is regulatory as well as technical: frameworks like the NIST AI Risk Management Framework and the EU AI Act are beginning to ask companies to demonstrate that their AI systems have been tested for exactly these failures, turning what used to be a nice-to-have into a compliance line item.

03 / WHAT IT DOESOne agent, every surface

Casco runs against four surfaces, and the breadth is part of the point. A single system probes the web app, the API layer, the cloud configuration, and the AI systems, rather than stitching together four separate vendors.

The product comes in two shapes. The first is fully autonomous testing that runs on its own and reports what it finds. The second, which the company calls supervised testing, keeps a human security expert in the loop to validate and extend the machine's work - useful for high-stakes systems or when an auditor wants a name attached. Both feed into compliance-ready reporting mapped to the frameworks companies actually get asked about: SOC 2, ISO 27001, the NIST AI Risk Management Framework, and the EU AI Act. The compliance document, in Casco's telling, is a byproduct of real testing rather than a binder assembled to satisfy a checkbox.

For the AI surface specifically, Casco runs continuous red-teaming - simulating the sophisticated, multi-step attacks a determined adversary would use against an agent, so the failure happens in a test instead of in production. The company also ships free public tools, including a CVSS calculator and a bug bounty calculator, which double as a front door to the security community it is courting.

04 / WHO USES ITStartups first, and the systems behind bigger names

The named customer list skews toward companies that ship fast and think about security early: the payroll platform Gusto, the AI-agent framework CrewAI, plus Scout, SixtyFour, Spreedly, Whop, Blaxel, Daytona, Archil, Novig, and Levels.fyi, among others. That mix is telling. These are teams building on the frontier - AI agents, developer infrastructure, fintech rails - where the attack surface is new enough that off-the-shelf scanners have not caught up. Matthew Broom, head of security at CrewAI, called Casco the first cybersecurity product to make security testing high-quality, fast, and easy at the same time.

05 / THE MARKETWhere a continuous attacker fits

Casco sits in a crowded field with a specific angle. On one side are the manual and boutique pentest firms and the pentest-as-a-service platforms - Cobalt, HackerOne, Bugcrowd, Synack - that coordinate human testers. On the other are automated application and API scanners that run fast but shallow. Casco's pitch is to collapse the trade-off: the depth of a skilled human tester, the speed and repeatability of a machine, and a fourth surface - AI systems - that most of the incumbents were not built to cover. The company goes as far as to claim its autonomous testing outperforms human penetration testers, a bold line it will spend the next few years being measured against.

The business model is straightforward B2B SaaS: a subscription for continuous testing, with supervised engagements and compliance reporting available on top. It is the kind of infrastructure that, done well, disappears into the background - the security equivalent of a smoke detector you forget about until the one night it matters.

06 / THE MONEYA year old, a nine-figure valuation

Casco went through Y Combinator's X25 batch under partner Jon Xu, and has since raised a reported Series A at a $100 million valuation, led by Standard Capital with participation from FundersClub, Y Combinator, and Rebel Fund. For a company founded in 2025, the trajectory is steep - customers, a nine-figure valuation, and a Gold Sponsorship of OWASP, the nonprofit that effectively writes the rulebook for application security. The sponsorship is a small detail with a large signal: Casco is planting a flag in the community whose standards it is building a product around.

What Casco is really selling is a change in tense. Security has always been described in the past - what got breached, what the audit found, what the report said. The company's whole design is to move that verb into the present continuous: something is testing your systems, right now, and it does not stop to file a report and go home. Whether the AI truly out-tests a seasoned human is the question the next few years will settle. But the framing has already done its work - once you have seen security testing run continuously, the annual ritual starts to look like what it is.